In central Pennsylvania, a shuttered nuclear reactor is being prepped for a comeback, a nearby operating plant is at the center of a first-of-its-kind federal dispute, and the common thread is the same: America’s largest technology companies want nuclear power for their artificial intelligence operations, and they want it now. As of spring 2026, Amazon, Microsoft, and Google have each staked out nuclear deals that collectively signal a structural shift in how the tech industry sources electricity, one that federal regulators are still scrambling to govern.
Why Big Tech turned to nuclear
AI data centers are not ordinary electricity consumers. Training and running large language models demands computing power that dwarfs traditional cloud workloads, and the facilities that house those chips run around the clock, every day of the year. A U.S. Department of Energy report, drawing on Lawrence Berkeley National Laboratory analysis, tracked data center electricity consumption from 2014 through 2023 and projected demand through 2028. The agency found that AI-driven growth is not a temporary spike but a system-level concern for grid reliability and affordability, with even modest percentage increases in national electricity use translating into enormous absolute loads when concentrated in a handful of high-density facilities.
That kind of demand profile fits nuclear power almost uniquely. Nuclear plants generate carbon-free electricity at massive scale, operate at capacity factors above 90 percent, and do not depend on weather. For tech companies under pressure from investors and regulators to meet climate commitments while powering exponentially growing AI workloads, reactors offer something wind and solar cannot: firm, 24/7 clean energy without the intermittency problem.
The deals taking shape
The most contentious arrangement so far involves Amazon Web Services and Talen Energy’s Susquehanna nuclear plant in Pennsylvania. Under a proposal filed with PJM Interconnection, the regional grid operator, AWS would co-locate a data center directly alongside the reactor and draw power behind the meter, bypassing the shared transmission grid entirely. The Federal Energy Regulatory Commission took up the case under docket ER24-2172, making it the first regulatory test of a behind-the-meter nuclear deal of this kind. In its order on the filing, FERC rejected the interconnection service agreement as proposed, citing concerns about reliability, cost allocation, and the precedent such an arrangement would set for wholesale power markets.
FERC Chairman Willie Phillips went further in a formal dissenting statement, calling reliability the commission’s “job number one.” Phillips argued that allowing a large corporate customer to claim power directly from a generation source raises novel legal and operational questions the commission had not adequately resolved. His central warning: if high-value nuclear output is redirected to a single buyer, the megawatts that previously supported the broader grid vanish from the shared pool, but the fixed costs of maintaining transmission lines, substations, and backup generation do not disappear with them. Remaining customers could end up paying more for less reliable service.
Meanwhile, Constellation Energy is pursuing the restart of Three Mile Island Unit 1, now designated the Christopher M. Crane Clean Energy Center. Microsoft signed a 20-year power purchase agreement with Constellation for the plant’s output, a deal that would channel carbon-free electricity to Microsoft’s data center operations. The Nuclear Regulatory Commission has established an information hub tracking the restart process, including licensing submissions and safety reviews. The NRC record confirms that federal oversight is active but that multiple approvals remain outstanding. No target completion date has been published, and any restart would need to satisfy current safety and environmental standards rather than simply reactivate a mothballed asset.
Google, for its part, has signed an agreement with Kairos Power to purchase electricity from small modular reactors, a next-generation nuclear technology that has not yet been deployed commercially in the United States. Amazon has also invested in X-energy, another SMR developer, broadening its nuclear bets beyond the Susquehanna co-location proposal. Taken together, these deals represent billions of dollars in committed or anticipated spending and mark the first time major tech companies have moved aggressively into nuclear procurement.
What regulators are worried about
The DOE report frames the problem in systemic terms. Each co-location deal or dedicated power purchase that removes generation capacity from the grid changes the supply picture for every other customer connected to that system. When a nuclear plant that once fed electricity into the shared network instead sends its output to a single corporate campus, grid planners must find replacement power, often from more expensive or less reliable sources. The cost of that substitution does not fall on the tech company behind the meter. It falls on households, hospitals, and small businesses still connected to the grid.
FERC’s handling of the Susquehanna case shows regulators are still developing the rules for these arrangements. The commission’s order tried to balance innovation in clean energy procurement against long-standing principles of open access and cost causation, the idea that those who impose costs on the grid should bear them. But the split decision and Phillips’s pointed dissent reveal deep disagreement about where that balance lies. With multiple similar proposals expected across PJM and other regional markets, the Susquehanna docket is less a final answer than a first draft.
State regulators and consumer advocates are watching closely but have limited tools. FERC governs wholesale markets and interstate transmission; state public utility commissions oversee retail rates and local distribution. A co-location deal structured to avoid the transmission grid can fall into a regulatory gap, visible to federal authorities but difficult for state officials to influence directly. No published institutional study yet quantifies how these deals would affect electricity prices for residential and commercial customers in PJM’s territory, leaving the cost question open.
The unknowns that will shape what comes next
Several critical variables remain unresolved. The exact financial terms of the AWS-Talen arrangement at Susquehanna have not been disclosed in accessible filings, so the true scale of Big Tech’s financial commitment to nuclear co-location is still opaque. The Three Mile Island restart timeline is open-ended: the NRC is processing submissions, but technical evaluations, public comment periods, and potential legal challenges could stretch the process for years. Grid operators and data center developers cannot reliably count on TMI-1 as a near-term source of firm power.
Even the DOE’s demand projections carry wide uncertainty bands. If AI adoption accelerates faster than current models assume, the upper end of the agency’s forecast could become the baseline, intensifying the scramble for dedicated power. If breakthroughs in chip efficiency or cooling technology slow electricity growth, the pressure to lock down exclusive nuclear supply could ease. Neither outcome is predictable with confidence based on available data.
What is clear is that the rules governing who gets to claim scarce, carbon-free electricity are being written right now. The Susquehanna docket, the Three Mile Island restart, and the emerging SMR deals are not isolated corporate transactions. They are test cases that will determine whether the nation’s nuclear fleet serves the grid as a whole or becomes a private power source for the companies wealthy enough to buy in. For the tech sector, the stakes are operational: without reliable clean power, AI ambitions hit a wall. For everyone else, the stakes are the monthly electric bill and the reliability of the lights when demand peaks.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
